static void
vc4_resource_transfer_unmap(struct pipe_context *pctx,
struct pipe_transfer *ptrans)
{
struct vc4_context *vc4 = vc4_context(pctx);
struct vc4_transfer *trans = vc4_transfer(ptrans);
struct pipe_resource *prsc = ptrans->resource;
struct vc4_resource *rsc = vc4_resource(prsc);
struct vc4_resource_slice *slice = &rsc->slices[ptrans->level];
if (trans->map) {
if (ptrans->usage & PIPE_TRANSFER_WRITE) {
vc4_store_tiled_image(rsc->bo->map + slice->offset +
ptrans->box.z * rsc->cube_map_stride,
slice->stride,
trans->map, ptrans->stride,
slice->tiling, rsc->cpp,
&ptrans->box);
}
free(trans->map);
}
pipe_resource_reference(&ptrans->resource, NULL);
util_slab_free(&vc4->transfer_pool, ptrans);
}
static void
vc4_submit_setup_rcl_render_config_surface(struct vc4_context *vc4,
struct drm_vc4_submit_rcl_surface *submit_surf,
struct pipe_surface *psurf)
{
struct vc4_surface *surf = vc4_surface(psurf);
if (!surf) {
submit_surf->hindex = ~0;
return;
}
struct vc4_resource *rsc = vc4_resource(psurf->texture);
submit_surf->hindex = vc4_gem_hindex(vc4, rsc->bo);
submit_surf->offset = surf->offset;
if (psurf->texture->nr_samples == 0) {
submit_surf->bits =
VC4_SET_FIELD(vc4_rt_format_is_565(surf->base.format) ?
VC4_RENDER_CONFIG_FORMAT_BGR565 :
VC4_RENDER_CONFIG_FORMAT_RGBA8888,
VC4_RENDER_CONFIG_FORMAT) |
VC4_SET_FIELD(surf->tiling,
VC4_RENDER_CONFIG_MEMORY_FORMAT);
}
rsc->writes++;
}
static struct pipe_surface *
vc4_create_surface(struct pipe_context *pctx,
struct pipe_resource *ptex,
const struct pipe_surface *surf_tmpl)
{
struct vc4_surface *surface = CALLOC_STRUCT(vc4_surface);
struct vc4_resource *rsc = vc4_resource(ptex);
if (!surface)
return NULL;
assert(surf_tmpl->u.tex.first_layer == surf_tmpl->u.tex.last_layer);
struct pipe_surface *psurf = &surface->base;
unsigned level = surf_tmpl->u.tex.level;
pipe_reference_init(&psurf->reference, 1);
pipe_resource_reference(&psurf->texture, ptex);
psurf->context = pctx;
psurf->format = surf_tmpl->format;
psurf->width = u_minify(ptex->width0, level);
psurf->height = u_minify(ptex->height0, level);
psurf->u.tex.level = level;
psurf->u.tex.first_layer = surf_tmpl->u.tex.first_layer;
psurf->u.tex.last_layer = surf_tmpl->u.tex.last_layer;
surface->offset = rsc->slices[level].offset;
surface->tiling = rsc->slices[level].tiling;
return &surface->base;
}
static void
vc4_set_index_buffer(struct pipe_context *pctx,
const struct pipe_index_buffer *ib)
{
struct vc4_context *vc4 = vc4_context(pctx);
if (ib) {
assert(!ib->user_buffer);
if (ib->index_size == 4) {
struct pipe_resource tmpl = *ib->buffer;
assert(tmpl.format == PIPE_FORMAT_R8_UNORM);
assert(tmpl.height0 == 1);
tmpl.width0 = (tmpl.width0 - ib->offset) / 2;
struct pipe_resource *pshadow =
vc4_resource_create(&vc4->screen->base, &tmpl);
struct vc4_resource *shadow = vc4_resource(pshadow);
pipe_resource_reference(&shadow->shadow_parent, ib->buffer);
pipe_resource_reference(&vc4->indexbuf.buffer, NULL);
vc4->indexbuf.buffer = pshadow;
vc4->indexbuf.index_size = 2;
vc4->indexbuf.offset = 0;
} else {
pipe_resource_reference(&vc4->indexbuf.buffer, ib->buffer);
vc4->indexbuf.index_size = ib->index_size;
vc4->indexbuf.offset = ib->offset;
}
} else {
pipe_resource_reference(&vc4->indexbuf.buffer, NULL);
}
vc4->dirty |= VC4_DIRTY_INDEXBUF;
}
static void
vc4_resource_destroy(struct pipe_screen *pscreen,
struct pipe_resource *prsc)
{
struct vc4_resource *rsc = vc4_resource(prsc);
pipe_resource_reference(&rsc->shadow_parent, NULL);
vc4_bo_unreference(&rsc->bo);
free(rsc);
}
static void
vc4_set_framebuffer_state(struct pipe_context *pctx,
const struct pipe_framebuffer_state *framebuffer)
{
struct vc4_context *vc4 = vc4_context(pctx);
struct pipe_framebuffer_state *cso = &vc4->framebuffer;
unsigned i;
vc4_flush(pctx);
for (i = 0; i < framebuffer->nr_cbufs; i++)
pipe_surface_reference(&cso->cbufs[i], framebuffer->cbufs[i]);
for (; i < vc4->framebuffer.nr_cbufs; i++)
pipe_surface_reference(&cso->cbufs[i], NULL);
cso->nr_cbufs = framebuffer->nr_cbufs;
pipe_surface_reference(&cso->zsbuf, framebuffer->zsbuf);
cso->width = framebuffer->width;
cso->height = framebuffer->height;
/* Nonzero texture mipmap levels are laid out as if they were in
* power-of-two-sized spaces. The renderbuffer config infers its
* stride from the width parameter, so we need to configure our
* framebuffer. Note that if the z/color buffers were mismatched
* sizes, we wouldn't be able to do this.
*/
if (cso->cbufs[0] && cso->cbufs[0]->u.tex.level) {
struct vc4_resource *rsc =
vc4_resource(cso->cbufs[0]->texture);
cso->width =
(rsc->slices[cso->cbufs[0]->u.tex.level].stride /
rsc->cpp);
} else if (cso->zsbuf && cso->zsbuf->u.tex.level){
struct vc4_resource *rsc =
vc4_resource(cso->zsbuf->texture);
cso->width =
(rsc->slices[cso->zsbuf->u.tex.level].stride /
rsc->cpp);
}
vc4->dirty |= VC4_DIRTY_FRAMEBUFFER;
}
static boolean
vc4_resource_get_handle(struct pipe_screen *pscreen,
struct pipe_resource *prsc,
struct winsys_handle *handle)
{
struct vc4_resource *rsc = vc4_resource(prsc);
return vc4_screen_bo_get_handle(pscreen, rsc->bo, rsc->slices[0].stride,
handle);
}
/**
* Flushes the current command lists if they reference the given BO.
*
* This helps avoid flushing the command buffers when unnecessary.
*/
bool
vc4_cl_references_bo(struct pipe_context *pctx, struct vc4_bo *bo,
bool include_reads)
{
struct vc4_context *vc4 = vc4_context(pctx);
if (!vc4->needs_flush)
return false;
/* Walk all the referenced BOs in the drawing command list to see if
* they match.
*/
if (include_reads) {
struct vc4_bo **referenced_bos = vc4->bo_pointers.base;
for (int i = 0; i < cl_offset(&vc4->bo_handles) / 4; i++) {
if (referenced_bos[i] == bo) {
return true;
}
}
}
/* Also check for the Z/color buffers, since the references to those
* are only added immediately before submit.
*/
struct vc4_surface *csurf = vc4_surface(vc4->framebuffer.cbufs[0]);
if (csurf) {
struct vc4_resource *ctex = vc4_resource(csurf->base.texture);
if (ctex->bo == bo) {
return true;
}
}
struct vc4_surface *zsurf = vc4_surface(vc4->framebuffer.zsbuf);
if (zsurf) {
struct vc4_resource *ztex =
vc4_resource(zsurf->base.texture);
if (ztex->bo == bo) {
return true;
}
}
return false;
}
static enum vc4_texture_data_type
get_resource_texture_format(struct pipe_resource *prsc)
{
struct vc4_resource *rsc = vc4_resource(prsc);
uint8_t format = vc4_get_tex_format(prsc->format);
if (!rsc->tiled) {
assert(format == VC4_TEXTURE_TYPE_RGBA8888);
return VC4_TEXTURE_TYPE_RGBA32R;
}
return format;
}
static void
vc4_update_shadow_textures(struct pipe_context *pctx,
struct vc4_texture_stateobj *stage_tex)
{
for (int i = 0; i < stage_tex->num_textures; i++) {
struct pipe_sampler_view *view = stage_tex->textures[i];
if (!view)
continue;
struct vc4_resource *rsc = vc4_resource(view->texture);
if (rsc->shadow_parent)
vc4_update_shadow_baselevel_texture(pctx, view);
}
}
static struct pipe_sampler_view *
vc4_create_sampler_view(struct pipe_context *pctx, struct pipe_resource *prsc,
const struct pipe_sampler_view *cso)
{
struct pipe_sampler_view *so = malloc(sizeof(*so));
if (!so)
return NULL;
*so = *cso;
pipe_reference(NULL, &prsc->reference);
/* There is no hardware level clamping, and the start address of a
* texture may be misaligned, so in that case we have to copy to a
* temporary.
*/
if (so->u.tex.first_level) {
struct vc4_resource *shadow_parent = vc4_resource(prsc);
struct pipe_resource tmpl = shadow_parent->base.b;
struct vc4_resource *clone;
tmpl.width0 = u_minify(tmpl.width0, so->u.tex.first_level);
tmpl.height0 = u_minify(tmpl.height0, so->u.tex.first_level);
tmpl.last_level = so->u.tex.last_level - so->u.tex.first_level;
prsc = vc4_resource_create(pctx->screen, &tmpl);
clone = vc4_resource(prsc);
clone->shadow_parent = &shadow_parent->base.b;
/* Flag it as needing update of the contents from the parent. */
clone->writes = shadow_parent->writes - 1;
}
so->texture = prsc;
so->reference.count = 1;
so->context = pctx;
return so;
}
static void
vc4_predraw_check_textures(struct pipe_context *pctx,
struct vc4_texture_stateobj *stage_tex)
{
struct vc4_context *vc4 = vc4_context(pctx);
for (int i = 0; i < stage_tex->num_textures; i++) {
struct pipe_sampler_view *view = stage_tex->textures[i];
if (!view)
continue;
struct vc4_resource *rsc = vc4_resource(view->texture);
if (rsc->shadow_parent)
vc4_update_shadow_baselevel_texture(pctx, view);
vc4_flush_jobs_writing_resource(vc4, view->texture);
}
}
static void
vc4_submit_setup_rcl_msaa_surface(struct vc4_context *vc4,
struct drm_vc4_submit_rcl_surface *submit_surf,
struct pipe_surface *psurf)
{
struct vc4_surface *surf = vc4_surface(psurf);
if (!surf) {
submit_surf->hindex = ~0;
return;
}
struct vc4_resource *rsc = vc4_resource(psurf->texture);
submit_surf->hindex = vc4_gem_hindex(vc4, rsc->bo);
submit_surf->offset = surf->offset;
submit_surf->bits = 0;
rsc->writes++;
}
static void
vc4_submit_setup_rcl_surface(struct vc4_context *vc4,
struct drm_vc4_submit_rcl_surface *submit_surf,
struct pipe_surface *psurf,
bool is_depth, bool is_write)
{
struct vc4_surface *surf = vc4_surface(psurf);
if (!surf) {
submit_surf->hindex = ~0;
return;
}
struct vc4_resource *rsc = vc4_resource(psurf->texture);
submit_surf->hindex = vc4_gem_hindex(vc4, rsc->bo);
submit_surf->offset = surf->offset;
if (psurf->texture->nr_samples == 0) {
if (is_depth) {
submit_surf->bits =
VC4_SET_FIELD(VC4_LOADSTORE_TILE_BUFFER_ZS,
VC4_LOADSTORE_TILE_BUFFER_BUFFER);
} else {
submit_surf->bits =
VC4_SET_FIELD(VC4_LOADSTORE_TILE_BUFFER_COLOR,
VC4_LOADSTORE_TILE_BUFFER_BUFFER) |
VC4_SET_FIELD(vc4_rt_format_is_565(psurf->format) ?
VC4_LOADSTORE_TILE_BUFFER_BGR565 :
VC4_LOADSTORE_TILE_BUFFER_RGBA8888,
VC4_LOADSTORE_TILE_BUFFER_FORMAT);
}
submit_surf->bits |=
VC4_SET_FIELD(surf->tiling,
VC4_LOADSTORE_TILE_BUFFER_TILING);
} else {
assert(!is_write);
submit_surf->flags |= VC4_SUBMIT_RCL_SURFACE_READ_IS_FULL_RES;
}
if (is_write)
rsc->writes++;
}
static void *
vc4_resource_transfer_map(struct pipe_context *pctx,
struct pipe_resource *prsc,
unsigned level, unsigned usage,
const struct pipe_box *box,
struct pipe_transfer **pptrans)
{
struct vc4_context *vc4 = vc4_context(pctx);
struct vc4_resource *rsc = vc4_resource(prsc);
struct vc4_resource_slice *slice = &rsc->slices[level];
struct vc4_transfer *trans;
struct pipe_transfer *ptrans;
enum pipe_format format = prsc->format;
char *buf;
if (usage & PIPE_TRANSFER_DISCARD_WHOLE_RESOURCE) {
vc4_resource_bo_alloc(rsc);
} else if (!(usage & PIPE_TRANSFER_UNSYNCHRONIZED)) {
if (vc4_cl_references_bo(pctx, rsc->bo)) {
if ((usage & PIPE_TRANSFER_DISCARD_RANGE) &&
prsc->last_level == 0 &&
prsc->width0 == box->width &&
prsc->height0 == box->height &&
prsc->depth0 == box->depth) {
vc4_resource_bo_alloc(rsc);
} else {
vc4_flush(pctx);
}
}
}
if (usage & PIPE_TRANSFER_WRITE)
rsc->writes++;
trans = util_slab_alloc(&vc4->transfer_pool);
if (!trans)
return NULL;
/* XXX: Handle DONTBLOCK, DISCARD_RANGE, PERSISTENT, COHERENT. */
/* util_slab_alloc() doesn't zero: */
memset(trans, 0, sizeof(*trans));
ptrans = &trans->base;
pipe_resource_reference(&ptrans->resource, prsc);
ptrans->level = level;
ptrans->usage = usage;
ptrans->box = *box;
/* Note that the current kernel implementation is synchronous, so no
* need to do syncing stuff here yet.
*/
buf = vc4_bo_map(rsc->bo);
if (!buf) {
fprintf(stderr, "Failed to map bo\n");
goto fail;
}
*pptrans = ptrans;
if (rsc->tiled) {
uint32_t utile_w = vc4_utile_width(rsc->cpp);
uint32_t utile_h = vc4_utile_height(rsc->cpp);
/* No direct mappings of tiled, since we need to manually
* tile/untile.
*/
if (usage & PIPE_TRANSFER_MAP_DIRECTLY)
return NULL;
/* We need to align the box to utile boundaries, since that's
* what load/store operate on.
*/
uint32_t box_start_x = ptrans->box.x & (utile_w - 1);
uint32_t box_start_y = ptrans->box.y & (utile_h - 1);
ptrans->box.width += box_start_x;
ptrans->box.x -= box_start_x;
ptrans->box.height += box_start_y;
ptrans->box.y -= box_start_y;
ptrans->box.width = align(ptrans->box.width, utile_w);
ptrans->box.height = align(ptrans->box.height, utile_h);
ptrans->stride = ptrans->box.width * rsc->cpp;
ptrans->layer_stride = ptrans->stride;
trans->map = malloc(ptrans->stride * ptrans->box.height);
if (usage & PIPE_TRANSFER_READ) {
vc4_load_tiled_image(trans->map, ptrans->stride,
buf + slice->offset +
box->z * rsc->cube_map_stride,
slice->stride,
slice->tiling, rsc->cpp,
&ptrans->box);
}
return (trans->map +
box_start_x * rsc->cpp +
box_start_y * ptrans->stride);
} else {
ptrans->stride = slice->stride;
ptrans->layer_stride = ptrans->stride;
return buf + slice->offset +
box->y / util_format_get_blockheight(format) * ptrans->stride +
box->x / util_format_get_blockwidth(format) * rsc->cpp +
box->z * rsc->cube_map_stride;
}
fail:
vc4_resource_transfer_unmap(pctx, ptrans);
return NULL;
}
static void
vc4_draw_vbo(struct pipe_context *pctx, const struct pipe_draw_info *info)
{
struct vc4_context *vc4 = vc4_context(pctx);
if (info->mode >= PIPE_PRIM_QUADS) {
util_primconvert_save_index_buffer(vc4->primconvert, &vc4->indexbuf);
util_primconvert_save_rasterizer_state(vc4->primconvert, &vc4->rasterizer->base);
util_primconvert_draw_vbo(vc4->primconvert, info);
perf_debug("Fallback conversion for %d %s vertices\n",
info->count, u_prim_name(info->mode));
return;
}
/* Before setting up the draw, do any fixup blits necessary. */
vc4_predraw_check_textures(pctx, &vc4->verttex);
vc4_predraw_check_textures(pctx, &vc4->fragtex);
vc4_hw_2116_workaround(pctx, info->count);
struct vc4_job *job = vc4_get_job_for_fbo(vc4);
vc4_get_draw_cl_space(job, info->count);
if (vc4->prim_mode != info->mode) {
vc4->prim_mode = info->mode;
vc4->dirty |= VC4_DIRTY_PRIM_MODE;
}
vc4_start_draw(vc4);
if (!vc4_update_compiled_shaders(vc4, info->mode)) {
debug_warn_once("shader compile failed, skipping draw call.\n");
return;
}
vc4_emit_state(pctx);
if ((vc4->dirty & (VC4_DIRTY_VTXBUF |
VC4_DIRTY_VTXSTATE |
VC4_DIRTY_PRIM_MODE |
VC4_DIRTY_RASTERIZER |
VC4_DIRTY_COMPILED_CS |
VC4_DIRTY_COMPILED_VS |
VC4_DIRTY_COMPILED_FS |
vc4->prog.cs->uniform_dirty_bits |
vc4->prog.vs->uniform_dirty_bits |
vc4->prog.fs->uniform_dirty_bits)) ||
vc4->last_index_bias != info->index_bias) {
vc4_emit_gl_shader_state(vc4, info, 0);
}
vc4->dirty = 0;
/* Note that the primitive type fields match with OpenGL/gallium
* definitions, up to but not including QUADS.
*/
struct vc4_cl_out *bcl = cl_start(&job->bcl);
if (info->indexed) {
uint32_t offset = vc4->indexbuf.offset;
uint32_t index_size = vc4->indexbuf.index_size;
struct pipe_resource *prsc;
if (vc4->indexbuf.index_size == 4) {
prsc = vc4_get_shadow_index_buffer(pctx, &vc4->indexbuf,
info->count, &offset);
index_size = 2;
} else {
if (vc4->indexbuf.user_buffer) {
prsc = NULL;
u_upload_data(vc4->uploader, 0,
info->count * index_size, 4,
vc4->indexbuf.user_buffer,
&offset, &prsc);
} else {
prsc = vc4->indexbuf.buffer;
}
}
struct vc4_resource *rsc = vc4_resource(prsc);
cl_start_reloc(&job->bcl, &bcl, 1);
cl_u8(&bcl, VC4_PACKET_GL_INDEXED_PRIMITIVE);
cl_u8(&bcl,
info->mode |
(index_size == 2 ?
VC4_INDEX_BUFFER_U16:
VC4_INDEX_BUFFER_U8));
cl_u32(&bcl, info->count);
cl_reloc(job, &job->bcl, &bcl, rsc->bo, offset);
cl_u32(&bcl, vc4->max_index);
job->draw_calls_queued++;
if (vc4->indexbuf.index_size == 4 || vc4->indexbuf.user_buffer)
pipe_resource_reference(&prsc, NULL);
} else {
uint32_t count = info->count;
uint32_t start = info->start;
uint32_t extra_index_bias = 0;
while (count) {
uint32_t this_count = count;
uint32_t step = count;
static const uint32_t max_verts = 65535;
/* GFXH-515 / SW-5891: The binner emits 16 bit indices
* for drawarrays, which means that if start + count >
* 64k it would truncate the top bits. Work around
* this by emitting a limited number of primitives at
* a time and reemitting the shader state pointing
* farther down the vertex attribute arrays.
*
* To do this properly for line loops or trifans, we'd
* need to make a new VB containing the first vertex
* plus whatever remainder.
*/
if (extra_index_bias) {
cl_end(&job->bcl, bcl);
vc4_emit_gl_shader_state(vc4, info,
extra_index_bias);
bcl = cl_start(&job->bcl);
}
if (start + count > max_verts) {
switch (info->mode) {
case PIPE_PRIM_POINTS:
this_count = step = max_verts;
break;
case PIPE_PRIM_LINES:
this_count = step = max_verts - (max_verts % 2);
break;
case PIPE_PRIM_LINE_STRIP:
this_count = max_verts;
step = max_verts - 1;
break;
case PIPE_PRIM_LINE_LOOP:
this_count = max_verts;
step = max_verts - 1;
debug_warn_once("unhandled line loop "
"looping behavior with "
">65535 verts\n");
break;
case PIPE_PRIM_TRIANGLES:
this_count = step = max_verts - (max_verts % 3);
break;
case PIPE_PRIM_TRIANGLE_STRIP:
this_count = max_verts;
step = max_verts - 2;
break;
default:
debug_warn_once("unhandled primitive "
"max vert count, truncating\n");
this_count = step = max_verts;
}
}
cl_u8(&bcl, VC4_PACKET_GL_ARRAY_PRIMITIVE);
cl_u8(&bcl, info->mode);
cl_u32(&bcl, this_count);
cl_u32(&bcl, start);
job->draw_calls_queued++;
count -= step;
extra_index_bias += start + step;
start = 0;
}
}
cl_end(&job->bcl, bcl);
/* We shouldn't have tripped the HW_2116 bug with the GFXH-515
* workaround.
*/
assert(job->draw_calls_queued <= VC4_HW_2116_COUNT);
if (vc4->zsa && vc4->framebuffer.zsbuf) {
struct vc4_resource *rsc =
vc4_resource(vc4->framebuffer.zsbuf->texture);
if (vc4->zsa->base.depth.enabled) {
job->resolve |= PIPE_CLEAR_DEPTH;
rsc->initialized_buffers = PIPE_CLEAR_DEPTH;
}
if (vc4->zsa->base.stencil[0].enabled) {
job->resolve |= PIPE_CLEAR_STENCIL;
rsc->initialized_buffers |= PIPE_CLEAR_STENCIL;
}
}
job->resolve |= PIPE_CLEAR_COLOR0;
if (vc4_debug & VC4_DEBUG_ALWAYS_FLUSH)
vc4_flush(pctx);
}
static void
vc4_setup_rcl(struct vc4_context *vc4)
{
struct vc4_surface *csurf = vc4_surface(vc4->framebuffer.cbufs[0]);
struct vc4_resource *ctex = csurf ? vc4_resource(csurf->base.texture) : NULL;
struct vc4_surface *zsurf = vc4_surface(vc4->framebuffer.zsbuf);
struct vc4_resource *ztex = zsurf ? vc4_resource(zsurf->base.texture) : NULL;
if (!csurf)
vc4->resolve &= ~PIPE_CLEAR_COLOR0;
if (!zsurf)
vc4->resolve &= ~(PIPE_CLEAR_DEPTH | PIPE_CLEAR_STENCIL);
uint32_t resolve_uncleared = vc4->resolve & ~vc4->cleared;
uint32_t width = vc4->framebuffer.width;
uint32_t height = vc4->framebuffer.height;
uint32_t stride_in_tiles = align(width, 64) / 64;
assert(vc4->draw_min_x != ~0 && vc4->draw_min_y != ~0);
uint32_t min_x_tile = vc4->draw_min_x / 64;
uint32_t min_y_tile = vc4->draw_min_y / 64;
uint32_t max_x_tile = (vc4->draw_max_x - 1) / 64;
uint32_t max_y_tile = (vc4->draw_max_y - 1) / 64;
uint32_t xtiles = max_x_tile - min_x_tile + 1;
uint32_t ytiles = max_y_tile - min_y_tile + 1;
#if 0
fprintf(stderr, "RCL: resolve 0x%x clear 0x%x resolve uncleared 0x%x\n",
vc4->resolve,
vc4->cleared,
resolve_uncleared);
#endif
uint32_t reloc_size = 9;
uint32_t clear_size = 14;
uint32_t config_size = 11 + reloc_size;
uint32_t loadstore_size = 7 + reloc_size;
uint32_t tilecoords_size = 3;
uint32_t branch_size = 5 + reloc_size;
uint32_t color_store_size = 1;
uint32_t semaphore_size = 1;
cl_ensure_space(&vc4->rcl,
clear_size +
config_size +
loadstore_size +
semaphore_size +
xtiles * ytiles * (loadstore_size * 4 +
tilecoords_size * 3 +
branch_size +
color_store_size));
if (vc4->cleared) {
cl_u8(&vc4->rcl, VC4_PACKET_CLEAR_COLORS);
cl_u32(&vc4->rcl, vc4->clear_color[0]);
cl_u32(&vc4->rcl, vc4->clear_color[1]);
cl_u32(&vc4->rcl, vc4->clear_depth);
cl_u8(&vc4->rcl, vc4->clear_stencil);
}
/* The rendering mode config determines the pointer that's used for
* VC4_PACKET_STORE_MS_TILE_BUFFER address computations. The kernel
* could handle a no-relocation rendering mode config and deny those
* packets, but instead we just tell the kernel we're doing our color
* rendering to the Z buffer, and just don't emit any of those
* packets.
*/
struct vc4_surface *render_surf = csurf ? csurf : zsurf;
struct vc4_resource *render_tex = vc4_resource(render_surf->base.texture);
cl_start_reloc(&vc4->rcl, 1);
cl_u8(&vc4->rcl, VC4_PACKET_TILE_RENDERING_MODE_CONFIG);
cl_reloc(vc4, &vc4->rcl, render_tex->bo, render_surf->offset);
cl_u16(&vc4->rcl, width);
cl_u16(&vc4->rcl, height);
cl_u16(&vc4->rcl, ((render_surf->tiling <<
VC4_RENDER_CONFIG_MEMORY_FORMAT_SHIFT) |
(vc4_rt_format_is_565(render_surf->base.format) ?
VC4_RENDER_CONFIG_FORMAT_BGR565 :
VC4_RENDER_CONFIG_FORMAT_RGBA8888)));
/* The tile buffer normally gets cleared when the previous tile is
* stored. If the clear values changed between frames, then the tile
* buffer has stale clear values in it, so we have to do a store in
* None mode (no writes) so that we trigger the tile buffer clear.
*
* Excess clearing is only a performance cost, since per-tile contents
* will be loaded/stored in the loop below.
*/
if (vc4->cleared & (PIPE_CLEAR_COLOR0 |
PIPE_CLEAR_DEPTH |
PIPE_CLEAR_STENCIL)) {
cl_u8(&vc4->rcl, VC4_PACKET_TILE_COORDINATES);
cl_u8(&vc4->rcl, 0);
cl_u8(&vc4->rcl, 0);
cl_u8(&vc4->rcl, VC4_PACKET_STORE_TILE_BUFFER_GENERAL);
cl_u16(&vc4->rcl, VC4_LOADSTORE_TILE_BUFFER_NONE);
cl_u32(&vc4->rcl, 0); /* no address, since we're in None mode */
}
uint32_t color_hindex = ctex ? vc4_gem_hindex(vc4, ctex->bo) : 0;
uint32_t depth_hindex = ztex ? vc4_gem_hindex(vc4, ztex->bo) : 0;
uint32_t tile_alloc_hindex = vc4_gem_hindex(vc4, vc4->tile_alloc);
for (int y = min_y_tile; y <= max_y_tile; y++) {
for (int x = min_x_tile; x <= max_x_tile; x++) {
bool end_of_frame = (x == max_x_tile &&
y == max_y_tile);
bool coords_emitted = false;
/* Note that the load doesn't actually occur until the
* tile coords packet is processed, and only one load
* may be outstanding at a time.
*/
if (resolve_uncleared & PIPE_CLEAR_COLOR) {
vc4_store_before_load(vc4, &coords_emitted);
cl_start_reloc(&vc4->rcl, 1);
cl_u8(&vc4->rcl, VC4_PACKET_LOAD_TILE_BUFFER_GENERAL);
cl_u8(&vc4->rcl,
VC4_LOADSTORE_TILE_BUFFER_COLOR |
(csurf->tiling <<
VC4_LOADSTORE_TILE_BUFFER_FORMAT_SHIFT));
cl_u8(&vc4->rcl,
vc4_rt_format_is_565(csurf->base.format) ?
VC4_LOADSTORE_TILE_BUFFER_BGR565 :
VC4_LOADSTORE_TILE_BUFFER_RGBA8888);
cl_reloc_hindex(&vc4->rcl, color_hindex,
csurf->offset);
vc4_tile_coordinates(vc4, x, y, &coords_emitted);
}
if (resolve_uncleared & (PIPE_CLEAR_DEPTH | PIPE_CLEAR_STENCIL)) {
vc4_store_before_load(vc4, &coords_emitted);
cl_start_reloc(&vc4->rcl, 1);
cl_u8(&vc4->rcl, VC4_PACKET_LOAD_TILE_BUFFER_GENERAL);
cl_u8(&vc4->rcl,
VC4_LOADSTORE_TILE_BUFFER_ZS |
(zsurf->tiling <<
VC4_LOADSTORE_TILE_BUFFER_FORMAT_SHIFT));
cl_u8(&vc4->rcl, 0);
cl_reloc_hindex(&vc4->rcl, depth_hindex,
zsurf->offset);
vc4_tile_coordinates(vc4, x, y, &coords_emitted);
}
/* Clipping depends on tile coordinates having been
* emitted, so make sure it's happened even if
* everything was cleared to start.
*/
vc4_tile_coordinates(vc4, x, y, &coords_emitted);
/* Wait for the binner before jumping to the first
* tile's lists.
*/
if (x == min_x_tile && y == min_y_tile)
cl_u8(&vc4->rcl, VC4_PACKET_WAIT_ON_SEMAPHORE);
cl_start_reloc(&vc4->rcl, 1);
cl_u8(&vc4->rcl, VC4_PACKET_BRANCH_TO_SUB_LIST);
cl_reloc_hindex(&vc4->rcl, tile_alloc_hindex,
(y * stride_in_tiles + x) * 32);
if (vc4->resolve & (PIPE_CLEAR_DEPTH | PIPE_CLEAR_STENCIL)) {
vc4_tile_coordinates(vc4, x, y, &coords_emitted);
cl_start_reloc(&vc4->rcl, 1);
cl_u8(&vc4->rcl, VC4_PACKET_STORE_TILE_BUFFER_GENERAL);
cl_u8(&vc4->rcl,
VC4_LOADSTORE_TILE_BUFFER_ZS |
(zsurf->tiling <<
VC4_LOADSTORE_TILE_BUFFER_FORMAT_SHIFT));
cl_u8(&vc4->rcl,
VC4_STORE_TILE_BUFFER_DISABLE_COLOR_CLEAR);
cl_reloc_hindex(&vc4->rcl, depth_hindex,
zsurf->offset |
((end_of_frame &&
!(vc4->resolve & PIPE_CLEAR_COLOR0)) ?
VC4_LOADSTORE_TILE_BUFFER_EOF : 0));
coords_emitted = false;
}
if (vc4->resolve & PIPE_CLEAR_COLOR0) {
vc4_tile_coordinates(vc4, x, y, &coords_emitted);
if (end_of_frame) {
cl_u8(&vc4->rcl,
VC4_PACKET_STORE_MS_TILE_BUFFER_AND_EOF);
} else {
cl_u8(&vc4->rcl,
VC4_PACKET_STORE_MS_TILE_BUFFER);
}
coords_emitted = false;
}
/* One of the bits needs to have been set that would
* have triggered an EOF.
*/
assert(vc4->resolve & (PIPE_CLEAR_COLOR0 |
PIPE_CLEAR_DEPTH |
PIPE_CLEAR_STENCIL));
/* Any coords emitted must also have been consumed by
* a store.
*/
assert(!coords_emitted);
}
}
if (vc4->resolve & PIPE_CLEAR_COLOR0)
ctex->writes++;
if (vc4->resolve & (PIPE_CLEAR_DEPTH | PIPE_CLEAR_STENCIL))
ztex->writes++;
}
static void
vc4_draw_vbo(struct pipe_context *pctx, const struct pipe_draw_info *info)
{
struct vc4_context *vc4 = vc4_context(pctx);
if (info->mode >= PIPE_PRIM_QUADS) {
util_primconvert_save_index_buffer(vc4->primconvert, &vc4->indexbuf);
util_primconvert_save_rasterizer_state(vc4->primconvert, &vc4->rasterizer->base);
util_primconvert_draw_vbo(vc4->primconvert, info);
perf_debug("Fallback conversion for %d %s vertices\n",
info->count, u_prim_name(info->mode));
return;
}
/* Before setting up the draw, do any fixup blits necessary. */
vc4_update_shadow_textures(pctx, &vc4->verttex);
vc4_update_shadow_textures(pctx, &vc4->fragtex);
vc4_hw_2116_workaround(pctx);
vc4_get_draw_cl_space(vc4);
if (vc4->prim_mode != info->mode) {
vc4->prim_mode = info->mode;
vc4->dirty |= VC4_DIRTY_PRIM_MODE;
}
vc4_start_draw(vc4);
vc4_update_compiled_shaders(vc4, info->mode);
vc4_emit_state(pctx);
if ((vc4->dirty & (VC4_DIRTY_VTXBUF |
VC4_DIRTY_VTXSTATE |
VC4_DIRTY_PRIM_MODE |
VC4_DIRTY_RASTERIZER |
VC4_DIRTY_COMPILED_CS |
VC4_DIRTY_COMPILED_VS |
VC4_DIRTY_COMPILED_FS |
vc4->prog.cs->uniform_dirty_bits |
vc4->prog.vs->uniform_dirty_bits |
vc4->prog.fs->uniform_dirty_bits)) ||
vc4->last_index_bias != info->index_bias) {
vc4_emit_gl_shader_state(vc4, info);
}
vc4->dirty = 0;
/* Note that the primitive type fields match with OpenGL/gallium
* definitions, up to but not including QUADS.
*/
struct vc4_cl_out *bcl = cl_start(&vc4->bcl);
if (info->indexed) {
uint32_t offset = vc4->indexbuf.offset;
uint32_t index_size = vc4->indexbuf.index_size;
struct pipe_resource *prsc;
if (vc4->indexbuf.index_size == 4) {
prsc = vc4_get_shadow_index_buffer(pctx, &vc4->indexbuf,
info->count, &offset);
index_size = 2;
} else {
if (vc4->indexbuf.user_buffer) {
prsc = NULL;
u_upload_data(vc4->uploader, 0,
info->count * index_size, 4,
vc4->indexbuf.user_buffer,
&offset, &prsc);
} else {
prsc = vc4->indexbuf.buffer;
}
}
struct vc4_resource *rsc = vc4_resource(prsc);
cl_start_reloc(&vc4->bcl, &bcl, 1);
cl_u8(&bcl, VC4_PACKET_GL_INDEXED_PRIMITIVE);
cl_u8(&bcl,
info->mode |
(index_size == 2 ?
VC4_INDEX_BUFFER_U16:
VC4_INDEX_BUFFER_U8));
cl_u32(&bcl, info->count);
cl_reloc(vc4, &vc4->bcl, &bcl, rsc->bo, offset);
cl_u32(&bcl, vc4->max_index);
if (vc4->indexbuf.index_size == 4 || vc4->indexbuf.user_buffer)
pipe_resource_reference(&prsc, NULL);
} else {
cl_u8(&bcl, VC4_PACKET_GL_ARRAY_PRIMITIVE);
cl_u8(&bcl, info->mode);
cl_u32(&bcl, info->count);
cl_u32(&bcl, info->start);
}
cl_end(&vc4->bcl, bcl);
if (vc4->zsa && vc4->zsa->base.depth.enabled) {
vc4->resolve |= PIPE_CLEAR_DEPTH;
}
if (vc4->zsa && vc4->zsa->base.stencil[0].enabled)
vc4->resolve |= PIPE_CLEAR_STENCIL;
vc4->resolve |= PIPE_CLEAR_COLOR0;
vc4->shader_rec_count++;
if (vc4_debug & VC4_DEBUG_ALWAYS_FLUSH)
vc4_flush(pctx);
}
static void
vc4_emit_gl_shader_state(struct vc4_context *vc4, const struct pipe_draw_info *info)
{
/* VC4_DIRTY_VTXSTATE */
struct vc4_vertex_stateobj *vtx = vc4->vtx;
/* VC4_DIRTY_VTXBUF */
struct vc4_vertexbuf_stateobj *vertexbuf = &vc4->vertexbuf;
/* The simulator throws a fit if VS or CS don't read an attribute, so
* we emit a dummy read.
*/
uint32_t num_elements_emit = MAX2(vtx->num_elements, 1);
/* Emit the shader record. */
struct vc4_cl_out *shader_rec =
cl_start_shader_reloc(&vc4->shader_rec, 3 + num_elements_emit);
/* VC4_DIRTY_PRIM_MODE | VC4_DIRTY_RASTERIZER */
cl_u16(&shader_rec,
VC4_SHADER_FLAG_ENABLE_CLIPPING |
VC4_SHADER_FLAG_FS_SINGLE_THREAD |
((info->mode == PIPE_PRIM_POINTS &&
vc4->rasterizer->base.point_size_per_vertex) ?
VC4_SHADER_FLAG_VS_POINT_SIZE : 0));
/* VC4_DIRTY_COMPILED_FS */
cl_u8(&shader_rec, 0); /* fs num uniforms (unused) */
cl_u8(&shader_rec, vc4->prog.fs->num_inputs);
cl_reloc(vc4, &vc4->shader_rec, &shader_rec, vc4->prog.fs->bo, 0);
cl_u32(&shader_rec, 0); /* UBO offset written by kernel */
/* VC4_DIRTY_COMPILED_VS */
cl_u16(&shader_rec, 0); /* vs num uniforms */
cl_u8(&shader_rec, vc4->prog.vs->vattrs_live);
cl_u8(&shader_rec, vc4->prog.vs->vattr_offsets[8]);
cl_reloc(vc4, &vc4->shader_rec, &shader_rec, vc4->prog.vs->bo, 0);
cl_u32(&shader_rec, 0); /* UBO offset written by kernel */
/* VC4_DIRTY_COMPILED_CS */
cl_u16(&shader_rec, 0); /* cs num uniforms */
cl_u8(&shader_rec, vc4->prog.cs->vattrs_live);
cl_u8(&shader_rec, vc4->prog.cs->vattr_offsets[8]);
cl_reloc(vc4, &vc4->shader_rec, &shader_rec, vc4->prog.cs->bo, 0);
cl_u32(&shader_rec, 0); /* UBO offset written by kernel */
uint32_t max_index = 0xffff;
for (int i = 0; i < vtx->num_elements; i++) {
struct pipe_vertex_element *elem = &vtx->pipe[i];
struct pipe_vertex_buffer *vb =
&vertexbuf->vb[elem->vertex_buffer_index];
struct vc4_resource *rsc = vc4_resource(vb->buffer);
/* not vc4->dirty tracked: vc4->last_index_bias */
uint32_t offset = (vb->buffer_offset +
elem->src_offset +
vb->stride * info->index_bias);
uint32_t vb_size = rsc->bo->size - offset;
uint32_t elem_size =
util_format_get_blocksize(elem->src_format);
cl_reloc(vc4, &vc4->shader_rec, &shader_rec, rsc->bo, offset);
cl_u8(&shader_rec, elem_size - 1);
cl_u8(&shader_rec, vb->stride);
cl_u8(&shader_rec, vc4->prog.vs->vattr_offsets[i]);
cl_u8(&shader_rec, vc4->prog.cs->vattr_offsets[i]);
if (vb->stride > 0) {
max_index = MIN2(max_index,
(vb_size - elem_size) / vb->stride);
}
}
if (vtx->num_elements == 0) {
assert(num_elements_emit == 1);
struct vc4_bo *bo = vc4_bo_alloc(vc4->screen, 4096, "scratch VBO");
cl_reloc(vc4, &vc4->shader_rec, &shader_rec, bo, 0);
cl_u8(&shader_rec, 16 - 1); /* element size */
cl_u8(&shader_rec, 0); /* stride */
cl_u8(&shader_rec, 0); /* VS VPM offset */
cl_u8(&shader_rec, 0); /* CS VPM offset */
vc4_bo_unreference(&bo);
}
cl_end(&vc4->shader_rec, shader_rec);
struct vc4_cl_out *bcl = cl_start(&vc4->bcl);
/* the actual draw call. */
cl_u8(&bcl, VC4_PACKET_GL_SHADER_STATE);
assert(vtx->num_elements <= 8);
/* Note that number of attributes == 0 in the packet means 8
* attributes. This field also contains the offset into shader_rec.
*/
cl_u32(&bcl, num_elements_emit & 0x7);
cl_end(&vc4->bcl, bcl);
vc4_write_uniforms(vc4, vc4->prog.fs,
&vc4->constbuf[PIPE_SHADER_FRAGMENT],
&vc4->fragtex);
vc4_write_uniforms(vc4, vc4->prog.vs,
&vc4->constbuf[PIPE_SHADER_VERTEX],
&vc4->verttex);
vc4_write_uniforms(vc4, vc4->prog.cs,
&vc4->constbuf[PIPE_SHADER_VERTEX],
&vc4->verttex);
vc4->last_index_bias = info->index_bias;
vc4->max_index = max_index;
}
static void
vc4_clear(struct pipe_context *pctx, unsigned buffers,
const union pipe_color_union *color, double depth, unsigned stencil)
{
struct vc4_context *vc4 = vc4_context(pctx);
struct vc4_job *job = vc4_get_job_for_fbo(vc4);
/* We can't flag new buffers for clearing once we've queued draws. We
* could avoid this by using the 3d engine to clear.
*/
if (job->draw_calls_queued) {
perf_debug("Flushing rendering to process new clear.\n");
vc4_job_submit(vc4, job);
job = vc4_get_job_for_fbo(vc4);
}
if (buffers & PIPE_CLEAR_COLOR0) {
struct vc4_resource *rsc =
vc4_resource(vc4->framebuffer.cbufs[0]->texture);
uint32_t clear_color;
if (vc4_rt_format_is_565(vc4->framebuffer.cbufs[0]->format)) {
/* In 565 mode, the hardware will be packing our color
* for us.
*/
clear_color = pack_rgba(PIPE_FORMAT_R8G8B8A8_UNORM,
color->f);
} else {
/* Otherwise, we need to do this packing because we
* support multiple swizzlings of RGBA8888.
*/
clear_color =
pack_rgba(vc4->framebuffer.cbufs[0]->format,
color->f);
}
job->clear_color[0] = job->clear_color[1] = clear_color;
rsc->initialized_buffers |= (buffers & PIPE_CLEAR_COLOR0);
}
if (buffers & PIPE_CLEAR_DEPTHSTENCIL) {
struct vc4_resource *rsc =
vc4_resource(vc4->framebuffer.zsbuf->texture);
unsigned zsclear = buffers & PIPE_CLEAR_DEPTHSTENCIL;
/* Clearing ZS will clear both Z and stencil, so if we're
* trying to clear just one then we need to draw a quad to do
* it instead.
*/
if ((zsclear == PIPE_CLEAR_DEPTH ||
zsclear == PIPE_CLEAR_STENCIL) &&
(rsc->initialized_buffers & ~(zsclear | job->cleared)) &&
util_format_is_depth_and_stencil(vc4->framebuffer.zsbuf->format)) {
perf_debug("Partial clear of Z+stencil buffer, "
"drawing a quad instead of fast clearing\n");
vc4_blitter_save(vc4);
util_blitter_clear(vc4->blitter,
vc4->framebuffer.width,
vc4->framebuffer.height,
1,
zsclear,
NULL, depth, stencil);
buffers &= ~zsclear;
if (!buffers)
return;
}
/* Though the depth buffer is stored with Z in the high 24,
* for this field we just need to store it in the low 24.
*/
if (buffers & PIPE_CLEAR_DEPTH) {
job->clear_depth = util_pack_z(PIPE_FORMAT_Z24X8_UNORM,
depth);
}
if (buffers & PIPE_CLEAR_STENCIL)
job->clear_stencil = stencil;
rsc->initialized_buffers |= zsclear;
}
job->draw_min_x = 0;
job->draw_min_y = 0;
job->draw_max_x = vc4->framebuffer.width;
job->draw_max_y = vc4->framebuffer.height;
job->cleared |= buffers;
job->resolve |= buffers;
vc4_start_draw(vc4);
}
int
vc4_simulator_flush(struct vc4_context *vc4, struct drm_vc4_submit_cl *args)
{
struct vc4_surface *csurf = vc4_surface(vc4->framebuffer.cbufs[0]);
struct vc4_resource *ctex = vc4_resource(csurf->base.texture);
uint32_t winsys_stride = ctex->bo->simulator_winsys_stride;
uint32_t sim_stride = ctex->slices[0].stride;
uint32_t row_len = MIN2(sim_stride, winsys_stride);
struct exec_info exec;
struct drm_device local_dev = {
.vc4 = vc4,
.simulator_mem_next = OVERFLOW_SIZE,
};
struct drm_device *dev = &local_dev;
int ret;
memset(&exec, 0, sizeof(exec));
if (ctex->bo->simulator_winsys_map) {
#if 0
fprintf(stderr, "%dx%d %d %d %d\n",
ctex->base.b.width0, ctex->base.b.height0,
winsys_stride,
sim_stride,
ctex->bo->size);
#endif
for (int y = 0; y < ctex->base.b.height0; y++) {
memcpy(ctex->bo->map + y * sim_stride,
ctex->bo->simulator_winsys_map + y * winsys_stride,
row_len);
}
}
exec.args = args;
ret = vc4_simulator_pin_bos(dev, &exec);
if (ret)
return ret;
ret = vc4_cl_validate(dev, &exec);
if (ret)
return ret;
simpenrose_do_binning(exec.ct0ca, exec.ct0ea);
simpenrose_do_rendering(exec.ct1ca, exec.ct1ea);
ret = vc4_simulator_unpin_bos(&exec);
if (ret)
return ret;
free(exec.exec_bo);
if (ctex->bo->simulator_winsys_map) {
for (int y = 0; y < ctex->base.b.height0; y++) {
memcpy(ctex->bo->simulator_winsys_map + y * winsys_stride,
ctex->bo->map + y * sim_stride,
row_len);
}
}
return 0;
}
static void
vc4_draw_vbo(struct pipe_context *pctx, const struct pipe_draw_info *info)
{
struct vc4_context *vc4 = vc4_context(pctx);
if (info->mode >= PIPE_PRIM_QUADS) {
util_primconvert_save_index_buffer(vc4->primconvert, &vc4->indexbuf);
util_primconvert_save_rasterizer_state(vc4->primconvert, &vc4->rasterizer->base);
util_primconvert_draw_vbo(vc4->primconvert, info);
return;
}
struct vc4_vertex_stateobj *vtx = vc4->vtx;
struct vc4_vertexbuf_stateobj *vertexbuf = &vc4->vertexbuf;
if (vc4->prim_mode != info->mode) {
vc4->prim_mode = info->mode;
vc4->dirty |= VC4_DIRTY_PRIM_MODE;
}
vc4_start_draw(vc4);
vc4_update_compiled_shaders(vc4, info->mode);
vc4_emit_state(pctx);
vc4->dirty = 0;
vc4_write_uniforms(vc4, vc4->prog.fs,
&vc4->constbuf[PIPE_SHADER_FRAGMENT],
&vc4->fragtex);
vc4_write_uniforms(vc4, vc4->prog.vs,
&vc4->constbuf[PIPE_SHADER_VERTEX],
&vc4->verttex);
vc4_write_uniforms(vc4, vc4->prog.cs,
&vc4->constbuf[PIPE_SHADER_VERTEX],
&vc4->verttex);
/* The simulator throws a fit if VS or CS don't read an attribute, so
* we emit a dummy read.
*/
uint32_t num_elements_emit = MAX2(vtx->num_elements, 1);
/* Emit the shader record. */
cl_start_shader_reloc(&vc4->shader_rec, 3 + num_elements_emit);
cl_u16(&vc4->shader_rec,
VC4_SHADER_FLAG_ENABLE_CLIPPING |
((info->mode == PIPE_PRIM_POINTS &&
vc4->rasterizer->base.point_size_per_vertex) ?
VC4_SHADER_FLAG_VS_POINT_SIZE : 0));
cl_u8(&vc4->shader_rec, 0); /* fs num uniforms (unused) */
cl_u8(&vc4->shader_rec, vc4->prog.fs->num_inputs);
cl_reloc(vc4, &vc4->shader_rec, vc4->prog.fs->bo, 0);
cl_u32(&vc4->shader_rec, 0); /* UBO offset written by kernel */
cl_u16(&vc4->shader_rec, 0); /* vs num uniforms */
cl_u8(&vc4->shader_rec, (1 << num_elements_emit) - 1); /* vs attribute array bitfield */
cl_u8(&vc4->shader_rec, 16 * num_elements_emit); /* vs total attribute size */
cl_reloc(vc4, &vc4->shader_rec, vc4->prog.vs->bo, 0);
cl_u32(&vc4->shader_rec, 0); /* UBO offset written by kernel */
cl_u16(&vc4->shader_rec, 0); /* cs num uniforms */
cl_u8(&vc4->shader_rec, (1 << num_elements_emit) - 1); /* cs attribute array bitfield */
cl_u8(&vc4->shader_rec, 16 * num_elements_emit); /* cs total attribute size */
cl_reloc(vc4, &vc4->shader_rec, vc4->prog.cs->bo, 0);
cl_u32(&vc4->shader_rec, 0); /* UBO offset written by kernel */
uint32_t max_index = 0xffff;
for (int i = 0; i < vtx->num_elements; i++) {
struct pipe_vertex_element *elem = &vtx->pipe[i];
struct pipe_vertex_buffer *vb =
&vertexbuf->vb[elem->vertex_buffer_index];
struct vc4_resource *rsc = vc4_resource(vb->buffer);
uint32_t offset = vb->buffer_offset + elem->src_offset;
uint32_t vb_size = rsc->bo->size - offset;
uint32_t elem_size =
util_format_get_blocksize(elem->src_format);
cl_reloc(vc4, &vc4->shader_rec, rsc->bo, offset);
cl_u8(&vc4->shader_rec, elem_size - 1);
cl_u8(&vc4->shader_rec, vb->stride);
cl_u8(&vc4->shader_rec, i * 16); /* VS VPM offset */
cl_u8(&vc4->shader_rec, i * 16); /* CS VPM offset */
if (vb->stride > 0) {
max_index = MIN2(max_index,
(vb_size - elem_size) / vb->stride);
}
}
if (vtx->num_elements == 0) {
assert(num_elements_emit == 1);
struct vc4_bo *bo = vc4_bo_alloc(vc4->screen, 4096, "scratch VBO");
cl_reloc(vc4, &vc4->shader_rec, bo, 0);
cl_u8(&vc4->shader_rec, 16 - 1); /* element size */
cl_u8(&vc4->shader_rec, 0); /* stride */
cl_u8(&vc4->shader_rec, 0); /* VS VPM offset */
cl_u8(&vc4->shader_rec, 0); /* CS VPM offset */
vc4_bo_unreference(&bo);
}
/* the actual draw call. */
cl_u8(&vc4->bcl, VC4_PACKET_GL_SHADER_STATE);
assert(vtx->num_elements <= 8);
/* Note that number of attributes == 0 in the packet means 8
* attributes. This field also contains the offset into shader_rec.
*/
cl_u32(&vc4->bcl, num_elements_emit & 0x7);
/* Note that the primitive type fields match with OpenGL/gallium
* definitions, up to but not including QUADS.
*/
if (info->indexed) {
struct vc4_resource *rsc = vc4_resource(vc4->indexbuf.buffer);
uint32_t offset = vc4->indexbuf.offset;
uint32_t index_size = vc4->indexbuf.index_size;
if (rsc->shadow_parent) {
vc4_update_shadow_index_buffer(pctx, &vc4->indexbuf);
offset = 0;
index_size = 2;
}
cl_start_reloc(&vc4->bcl, 1);
cl_u8(&vc4->bcl, VC4_PACKET_GL_INDEXED_PRIMITIVE);
cl_u8(&vc4->bcl,
info->mode |
(index_size == 2 ?
VC4_INDEX_BUFFER_U16:
VC4_INDEX_BUFFER_U8));
cl_u32(&vc4->bcl, info->count);
cl_reloc(vc4, &vc4->bcl, rsc->bo, offset);
cl_u32(&vc4->bcl, max_index);
} else {
cl_u8(&vc4->bcl, VC4_PACKET_GL_ARRAY_PRIMITIVE);
cl_u8(&vc4->bcl, info->mode);
cl_u32(&vc4->bcl, info->count);
cl_u32(&vc4->bcl, info->start);
}
if (vc4->zsa && vc4->zsa->base.depth.enabled) {
vc4->resolve |= PIPE_CLEAR_DEPTH;
}
if (vc4->zsa && vc4->zsa->base.stencil[0].enabled)
vc4->resolve |= PIPE_CLEAR_STENCIL;
vc4->resolve |= PIPE_CLEAR_COLOR0;
vc4->shader_rec_count++;
if (vc4_debug & VC4_DEBUG_ALWAYS_FLUSH)
vc4_flush(pctx);
}
void
vc4_update_shadow_baselevel_texture(struct pipe_context *pctx,
struct pipe_sampler_view *view)
{
struct vc4_resource *shadow = vc4_resource(view->texture);
struct vc4_resource *orig = vc4_resource(shadow->shadow_parent);
assert(orig);
if (shadow->writes == orig->writes)
return;
for (int i = 0; i <= shadow->base.b.last_level; i++) {
struct pipe_box box = {
.x = 0,
.y = 0,
.z = 0,
.width = u_minify(shadow->base.b.width0, i),
.height = u_minify(shadow->base.b.height0, i),
.depth = 1,
};
util_resource_copy_region(pctx,
&shadow->base.b, i, 0, 0, 0,
&orig->base.b,
view->u.tex.first_level + i,
&box);
}
shadow->writes = orig->writes;
}
/**
* Converts a 4-byte index buffer to 2 bytes.
*
* Since GLES2 only has support for 1 and 2-byte indices, the hardware doesn't
* include 4-byte index support, and we have to shrink it down.
*
* There's no fallback support for when indices end up being larger than 2^16,
* though it will at least assertion fail. Also, if the original index data
* was in user memory, it would be nice to not have uploaded it to a VBO
* before translating.
*/
void
vc4_update_shadow_index_buffer(struct pipe_context *pctx,
const struct pipe_index_buffer *ib)
{
struct vc4_resource *shadow = vc4_resource(ib->buffer);
struct vc4_resource *orig = vc4_resource(shadow->shadow_parent);
uint32_t count = shadow->base.b.width0 / 2;
if (shadow->writes == orig->writes)
return;
struct pipe_transfer *src_transfer;
uint32_t *src = pipe_buffer_map_range(pctx, &orig->base.b,
ib->offset,
count * 4,
PIPE_TRANSFER_READ, &src_transfer);
struct pipe_transfer *dst_transfer;
uint16_t *dst = pipe_buffer_map_range(pctx, &shadow->base.b,
0,
count * 2,
PIPE_TRANSFER_WRITE, &dst_transfer);
for (int i = 0; i < count; i++) {
uint32_t src_index = src[i];
assert(src_index <= 0xffff);
dst[i] = src_index;
}
pctx->transfer_unmap(pctx, dst_transfer);
pctx->transfer_unmap(pctx, src_transfer);
shadow->writes = orig->writes;
}
void
vc4_resource_screen_init(struct pipe_screen *pscreen)
{
pscreen->resource_create = vc4_resource_create;
pscreen->resource_from_handle = vc4_resource_from_handle;
pscreen->resource_get_handle = u_resource_get_handle_vtbl;
pscreen->resource_destroy = u_resource_destroy_vtbl;
}
static void
vc4_set_framebuffer_state(struct pipe_context *pctx,
const struct pipe_framebuffer_state *framebuffer)
{
struct vc4_context *vc4 = vc4_context(pctx);
struct pipe_framebuffer_state *cso = &vc4->framebuffer;
unsigned i;
vc4_flush(pctx);
for (i = 0; i < framebuffer->nr_cbufs; i++)
pipe_surface_reference(&cso->cbufs[i], framebuffer->cbufs[i]);
for (; i < vc4->framebuffer.nr_cbufs; i++)
pipe_surface_reference(&cso->cbufs[i], NULL);
cso->nr_cbufs = framebuffer->nr_cbufs;
pipe_surface_reference(&cso->zsbuf, framebuffer->zsbuf);
cso->width = framebuffer->width;
cso->height = framebuffer->height;
/* If we're binding to uninitialized buffers, no need to load their
* contents before drawing..
*/
if (cso->cbufs[0]) {
struct vc4_resource *rsc =
vc4_resource(cso->cbufs[0]->texture);
if (!rsc->writes)
vc4->cleared |= PIPE_CLEAR_COLOR0;
}
if (cso->zsbuf) {
struct vc4_resource *rsc =
vc4_resource(cso->zsbuf->texture);
if (!rsc->writes)
vc4->cleared |= PIPE_CLEAR_DEPTH | PIPE_CLEAR_STENCIL;
}
/* Nonzero texture mipmap levels are laid out as if they were in
* power-of-two-sized spaces. The renderbuffer config infers its
* stride from the width parameter, so we need to configure our
* framebuffer. Note that if the z/color buffers were mismatched
* sizes, we wouldn't be able to do this.
*/
if (cso->cbufs[0] && cso->cbufs[0]->u.tex.level) {
struct vc4_resource *rsc =
vc4_resource(cso->cbufs[0]->texture);
cso->width =
(rsc->slices[cso->cbufs[0]->u.tex.level].stride /
rsc->cpp);
} else if (cso->zsbuf && cso->zsbuf->u.tex.level){
struct vc4_resource *rsc =
vc4_resource(cso->zsbuf->texture);
cso->width =
(rsc->slices[cso->zsbuf->u.tex.level].stride /
rsc->cpp);
}
vc4->msaa = false;
if (cso->cbufs[0])
vc4->msaa = cso->cbufs[0]->texture->nr_samples > 1;
else if (cso->zsbuf)
vc4->msaa = cso->zsbuf->texture->nr_samples > 1;
if (vc4->msaa) {
vc4->tile_width = 32;
vc4->tile_height = 32;
} else {
vc4->tile_width = 64;
vc4->tile_height = 64;
}
vc4->draw_tiles_x = DIV_ROUND_UP(cso->width, vc4->tile_width);
vc4->draw_tiles_y = DIV_ROUND_UP(cso->height, vc4->tile_height);
vc4->dirty |= VC4_DIRTY_FRAMEBUFFER;
}
int
vc4_simulator_flush(struct vc4_context *vc4, struct drm_vc4_submit_cl *args)
{
struct vc4_screen *screen = vc4->screen;
struct vc4_surface *csurf = vc4_surface(vc4->framebuffer.cbufs[0]);
struct vc4_resource *ctex = csurf ? vc4_resource(csurf->base.texture) : NULL;
uint32_t winsys_stride = ctex ? ctex->bo->simulator_winsys_stride : 0;
uint32_t sim_stride = ctex ? ctex->slices[0].stride : 0;
uint32_t row_len = MIN2(sim_stride, winsys_stride);
struct exec_info exec;
struct drm_device local_dev = {
.vc4 = vc4,
.simulator_mem_next = OVERFLOW_SIZE,
};
struct drm_device *dev = &local_dev;
int ret;
memset(&exec, 0, sizeof(exec));
if (ctex && ctex->bo->simulator_winsys_map) {
#if 0
fprintf(stderr, "%dx%d %d %d %d\n",
ctex->base.b.width0, ctex->base.b.height0,
winsys_stride,
sim_stride,
ctex->bo->size);
#endif
for (int y = 0; y < ctex->base.b.height0; y++) {
memcpy(ctex->bo->map + y * sim_stride,
ctex->bo->simulator_winsys_map + y * winsys_stride,
row_len);
}
}
exec.args = args;
ret = vc4_simulator_pin_bos(dev, &exec);
if (ret)
return ret;
ret = vc4_cl_validate(dev, &exec);
if (ret)
return ret;
int bfc = simpenrose_do_binning(exec.ct0ca, exec.ct0ea);
if (bfc != 1) {
fprintf(stderr, "Binning returned %d flushes, should be 1.\n",
bfc);
fprintf(stderr, "Relocated binning command list:\n");
vc4_dump_cl(screen->simulator_mem_base + exec.ct0ca,
exec.ct0ea - exec.ct0ca, false);
abort();
}
int rfc = simpenrose_do_rendering(exec.ct1ca, exec.ct1ea);
if (rfc != 1) {
fprintf(stderr, "Rendering returned %d frames, should be 1.\n",
rfc);
fprintf(stderr, "Relocated render command list:\n");
vc4_dump_cl(screen->simulator_mem_base + exec.ct1ca,
exec.ct1ea - exec.ct1ca, true);
abort();
}
ret = vc4_simulator_unpin_bos(&exec);
if (ret)
return ret;
vc4_bo_unreference(&exec.exec_bo->bo);
free(exec.exec_bo);
if (ctex && ctex->bo->simulator_winsys_map) {
for (int y = 0; y < ctex->base.b.height0; y++) {
memcpy(ctex->bo->simulator_winsys_map + y * winsys_stride,
ctex->bo->map + y * sim_stride,
row_len);
}
}
return 0;
}
static void *vc4_get_yuv_vs(struct pipe_context *pctx)
{
struct vc4_context *vc4 = vc4_context(pctx);
struct pipe_screen *pscreen = pctx->screen;
if (vc4->yuv_linear_blit_vs)
return vc4->yuv_linear_blit_vs;
const struct nir_shader_compiler_options *options =
pscreen->get_compiler_options(pscreen,
PIPE_SHADER_IR_NIR,
PIPE_SHADER_VERTEX);
nir_builder b;
nir_builder_init_simple_shader(&b, NULL, MESA_SHADER_VERTEX, options);
b.shader->info.name = ralloc_strdup(b.shader, "linear_blit_vs");
const struct glsl_type *vec4 = glsl_vec4_type();
nir_variable *pos_in = nir_variable_create(b.shader, nir_var_shader_in,
vec4, "pos");
nir_variable *pos_out = nir_variable_create(b.shader, nir_var_shader_out,
vec4, "gl_Position");
pos_out->data.location = VARYING_SLOT_POS;
nir_store_var(&b, pos_out, nir_load_var(&b, pos_in), 0xf);
struct pipe_shader_state shader_tmpl = {
.type = PIPE_SHADER_IR_NIR,
.ir.nir = b.shader,
};
vc4->yuv_linear_blit_vs = pctx->create_vs_state(pctx, &shader_tmpl);
return vc4->yuv_linear_blit_vs;
}
static void *vc4_get_yuv_fs(struct pipe_context *pctx, int cpp)
{
struct vc4_context *vc4 = vc4_context(pctx);
struct pipe_screen *pscreen = pctx->screen;
struct pipe_shader_state **cached_shader;
const char *name;
if (cpp == 1) {
cached_shader = &vc4->yuv_linear_blit_fs_8bit;
name = "linear_blit_8bit_fs";
} else {
cached_shader = &vc4->yuv_linear_blit_fs_16bit;
name = "linear_blit_16bit_fs";
}
if (*cached_shader)
return *cached_shader;
const struct nir_shader_compiler_options *options =
pscreen->get_compiler_options(pscreen,
PIPE_SHADER_IR_NIR,
PIPE_SHADER_FRAGMENT);
nir_builder b;
nir_builder_init_simple_shader(&b, NULL, MESA_SHADER_FRAGMENT, options);
b.shader->info.name = ralloc_strdup(b.shader, name);
const struct glsl_type *vec4 = glsl_vec4_type();
const struct glsl_type *glsl_int = glsl_int_type();
nir_variable *color_out = nir_variable_create(b.shader, nir_var_shader_out,
vec4, "f_color");
color_out->data.location = FRAG_RESULT_COLOR;
nir_variable *pos_in = nir_variable_create(b.shader, nir_var_shader_in,
vec4, "pos");
pos_in->data.location = VARYING_SLOT_POS;
nir_ssa_def *pos = nir_load_var(&b, pos_in);
nir_ssa_def *one = nir_imm_int(&b, 1);
nir_ssa_def *two = nir_imm_int(&b, 2);
nir_ssa_def *x = nir_f2i32(&b, nir_channel(&b, pos, 0));
nir_ssa_def *y = nir_f2i32(&b, nir_channel(&b, pos, 1));
nir_variable *stride_in = nir_variable_create(b.shader, nir_var_uniform,
glsl_int, "stride");
nir_ssa_def *stride = nir_load_var(&b, stride_in);
nir_ssa_def *x_offset;
nir_ssa_def *y_offset;
if (cpp == 1) {
nir_ssa_def *intra_utile_x_offset =
nir_ishl(&b, nir_iand(&b, x, one), two);
nir_ssa_def *inter_utile_x_offset =
nir_ishl(&b, nir_iand(&b, x, nir_imm_int(&b, ~3)), one);
x_offset = nir_iadd(&b,
intra_utile_x_offset,
inter_utile_x_offset);
y_offset = nir_imul(&b,
nir_iadd(&b,
nir_ishl(&b, y, one),
nir_ushr(&b, nir_iand(&b, x, two), one)),
stride);
} else {
x_offset = nir_ishl(&b, x, two);
y_offset = nir_imul(&b, y, stride);
}
nir_intrinsic_instr *load =
nir_intrinsic_instr_create(b.shader, nir_intrinsic_load_ubo);
load->num_components = 1;
nir_ssa_dest_init(&load->instr, &load->dest, load->num_components, 32, NULL);
load->src[0] = nir_src_for_ssa(one);
load->src[1] = nir_src_for_ssa(nir_iadd(&b, x_offset, y_offset));
nir_builder_instr_insert(&b, &load->instr);
nir_store_var(&b, color_out,
nir_unpack_unorm_4x8(&b, &load->dest.ssa),
0xf);
struct pipe_shader_state shader_tmpl = {
.type = PIPE_SHADER_IR_NIR,
.ir.nir = b.shader,
};
*cached_shader = pctx->create_fs_state(pctx, &shader_tmpl);
return *cached_shader;
}
static bool
vc4_yuv_blit(struct pipe_context *pctx, const struct pipe_blit_info *info)
{
struct vc4_context *vc4 = vc4_context(pctx);
struct vc4_resource *src = vc4_resource(info->src.resource);
struct vc4_resource *dst = vc4_resource(info->dst.resource);
bool ok;
if (src->tiled)
return false;
if (src->base.format != PIPE_FORMAT_R8_UNORM &&
src->base.format != PIPE_FORMAT_R8G8_UNORM)
return false;
/* YUV blits always turn raster-order to tiled */
assert(dst->base.format == src->base.format);
assert(dst->tiled);
/* Always 1:1 and at the origin */
assert(info->src.box.x == 0 && info->dst.box.x == 0);
assert(info->src.box.y == 0 && info->dst.box.y == 0);
assert(info->src.box.width == info->dst.box.width);
assert(info->src.box.height == info->dst.box.height);
if ((src->slices[info->src.level].offset & 3) ||
(src->slices[info->src.level].stride & 3)) {
perf_debug("YUV-blit src texture offset/stride misaligned: 0x%08x/%d\n",
src->slices[info->src.level].offset,
src->slices[info->src.level].stride);
goto fallback;
}
vc4_blitter_save(vc4);
/* Create a renderable surface mapping the T-tiled shadow buffer.
*/
struct pipe_surface dst_tmpl;
util_blitter_default_dst_texture(&dst_tmpl, info->dst.resource,
info->dst.level, info->dst.box.z);
dst_tmpl.format = PIPE_FORMAT_RGBA8888_UNORM;
struct pipe_surface *dst_surf =
pctx->create_surface(pctx, info->dst.resource, &dst_tmpl);
if (!dst_surf) {
fprintf(stderr, "Failed to create YUV dst surface\n");
util_blitter_unset_running_flag(vc4->blitter);
return false;
}
dst_surf->width /= 2;
if (dst->cpp == 1)
dst_surf->height /= 2;
/* Set the constant buffer. */
uint32_t stride = src->slices[info->src.level].stride;
struct pipe_constant_buffer cb_uniforms = {
.user_buffer = &stride,
.buffer_size = sizeof(stride),
};
pctx->set_constant_buffer(pctx, PIPE_SHADER_FRAGMENT, 0, &cb_uniforms);
struct pipe_constant_buffer cb_src = {
.buffer = info->src.resource,
.buffer_offset = src->slices[info->src.level].offset,
.buffer_size = (src->bo->size -
src->slices[info->src.level].offset),
};
pctx->set_constant_buffer(pctx, PIPE_SHADER_FRAGMENT, 1, &cb_src);
/* Unbind the textures, to make sure we don't try to recurse into the
* shadow blit.
*/
pctx->set_sampler_views(pctx, PIPE_SHADER_FRAGMENT, 0, 0, NULL);
pctx->bind_sampler_states(pctx, PIPE_SHADER_FRAGMENT, 0, 0, NULL);
util_blitter_custom_shader(vc4->blitter, dst_surf,
vc4_get_yuv_vs(pctx),
vc4_get_yuv_fs(pctx, src->cpp));
util_blitter_restore_textures(vc4->blitter);
util_blitter_restore_constant_buffer_state(vc4->blitter);
/* Restore cb1 (util_blitter doesn't handle this one). */
struct pipe_constant_buffer cb_disabled = { 0 };
pctx->set_constant_buffer(pctx, PIPE_SHADER_FRAGMENT, 1, &cb_disabled);
pipe_surface_reference(&dst_surf, NULL);
return true;
fallback:
/* Do an immediate SW fallback, since the render blit path
* would just recurse.
*/
ok = util_try_blit_via_copy_region(pctx, info);
assert(ok); (void)ok;
return true;
}
static bool
vc4_render_blit(struct pipe_context *ctx, struct pipe_blit_info *info)
{
struct vc4_context *vc4 = vc4_context(ctx);
if (!util_blitter_is_blit_supported(vc4->blitter, info)) {
fprintf(stderr, "blit unsupported %s -> %s\n",
util_format_short_name(info->src.resource->format),
util_format_short_name(info->dst.resource->format));
return false;
}
/* Enable the scissor, so we get a minimal set of tiles rendered. */
if (!info->scissor_enable) {
info->scissor_enable = true;
info->scissor.minx = info->dst.box.x;
info->scissor.miny = info->dst.box.y;
info->scissor.maxx = info->dst.box.x + info->dst.box.width;
info->scissor.maxy = info->dst.box.y + info->dst.box.height;
}
vc4_blitter_save(vc4);
util_blitter_blit(vc4->blitter, info);
return true;
}
/* Optimal hardware path for blitting pixels.
* Scaling, format conversion, up- and downsampling (resolve) are allowed.
*/
void
vc4_blit(struct pipe_context *pctx, const struct pipe_blit_info *blit_info)
{
struct pipe_blit_info info = *blit_info;
if (vc4_yuv_blit(pctx, blit_info))
return;
if (vc4_tile_blit(pctx, blit_info))
return;
if (info.mask & PIPE_MASK_S) {
if (util_try_blit_via_copy_region(pctx, &info))
return;
info.mask &= ~PIPE_MASK_S;
fprintf(stderr, "cannot blit stencil, skipping\n");
}
if (vc4_render_blit(pctx, &info))
return;
fprintf(stderr, "Unsupported blit\n");
}
static bool
vc4_tile_blit(struct pipe_context *pctx, const struct pipe_blit_info *info)
{
struct vc4_context *vc4 = vc4_context(pctx);
bool msaa = (info->src.resource->nr_samples > 1 ||
info->dst.resource->nr_samples > 1);
int tile_width = msaa ? 32 : 64;
int tile_height = msaa ? 32 : 64;
if (util_format_is_depth_or_stencil(info->dst.resource->format))
return false;
if (info->scissor_enable)
return false;
if ((info->mask & PIPE_MASK_RGBA) == 0)
return false;
if (info->dst.box.x != info->src.box.x ||
info->dst.box.y != info->src.box.y ||
info->dst.box.width != info->src.box.width ||
info->dst.box.height != info->src.box.height) {
return false;
}
int dst_surface_width = u_minify(info->dst.resource->width0,
info->dst.level);
int dst_surface_height = u_minify(info->dst.resource->height0,
info->dst.level);
if (is_tile_unaligned(info->dst.box.x, tile_width) ||
is_tile_unaligned(info->dst.box.y, tile_height) ||
(is_tile_unaligned(info->dst.box.width, tile_width) &&
info->dst.box.x + info->dst.box.width != dst_surface_width) ||
(is_tile_unaligned(info->dst.box.height, tile_height) &&
info->dst.box.y + info->dst.box.height != dst_surface_height)) {
return false;
}
/* VC4_PACKET_LOAD_TILE_BUFFER_GENERAL uses the
* VC4_PACKET_TILE_RENDERING_MODE_CONFIG's width (determined by our
* destination surface) to determine the stride. This may be wrong
* when reading from texture miplevels > 0, which are stored in
* POT-sized areas. For MSAA, the tile addresses are computed
* explicitly by the RCL, but still use the destination width to
* determine the stride (which could be fixed by explicitly supplying
* it in the ABI).
*/
struct vc4_resource *rsc = vc4_resource(info->src.resource);
uint32_t stride;
if (info->src.resource->nr_samples > 1)
stride = align(dst_surface_width, 32) * 4 * rsc->cpp;
else if (rsc->slices[info->src.level].tiling == VC4_TILING_FORMAT_T)
stride = align(dst_surface_width * rsc->cpp, 128);
else
stride = align(dst_surface_width * rsc->cpp, 16);
if (stride != rsc->slices[info->src.level].stride)
return false;
if (info->dst.resource->format != info->src.resource->format)
return false;
if (false) {
fprintf(stderr, "RCL blit from %d,%d to %d,%d (%d,%d)\n",
info->src.box.x,
info->src.box.y,
info->dst.box.x,
info->dst.box.y,
info->dst.box.width,
info->dst.box.height);
}
struct pipe_surface *dst_surf =
vc4_get_blit_surface(pctx, info->dst.resource, info->dst.level);
struct pipe_surface *src_surf =
vc4_get_blit_surface(pctx, info->src.resource, info->src.level);
vc4_flush_jobs_reading_resource(vc4, info->src.resource);
struct vc4_job *job = vc4_get_job(vc4, dst_surf, NULL);
pipe_surface_reference(&job->color_read, src_surf);
/* If we're resolving from MSAA to single sample, we still need to run
* the engine in MSAA mode for the load.
*/
if (!job->msaa && info->src.resource->nr_samples > 1) {
job->msaa = true;
job->tile_width = 32;
job->tile_height = 32;
}
job->draw_min_x = info->dst.box.x;
job->draw_min_y = info->dst.box.y;
job->draw_max_x = info->dst.box.x + info->dst.box.width;
job->draw_max_y = info->dst.box.y + info->dst.box.height;
job->draw_width = dst_surf->width;
job->draw_height = dst_surf->height;
job->tile_width = tile_width;
job->tile_height = tile_height;
job->msaa = msaa;
job->needs_flush = true;
job->resolve |= PIPE_CLEAR_COLOR;
vc4_job_submit(vc4, job);
pipe_surface_reference(&dst_surf, NULL);
pipe_surface_reference(&src_surf, NULL);
return true;
}
static void
vc4_setup_rcl(struct vc4_context *vc4)
{
struct vc4_surface *csurf = vc4_surface(vc4->framebuffer.cbufs[0]);
struct vc4_resource *ctex = vc4_resource(csurf->base.texture);
uint32_t resolve_uncleared = vc4->resolve & ~vc4->cleared;
uint32_t width = vc4->framebuffer.width;
uint32_t height = vc4->framebuffer.height;
uint32_t xtiles = align(width, 64) / 64;
uint32_t ytiles = align(height, 64) / 64;
#if 0
fprintf(stderr, "RCL: resolve 0x%x clear 0x%x resolve uncleared 0x%x\n",
vc4->resolve,
vc4->cleared,
resolve_uncleared);
#endif
cl_u8(&vc4->rcl, VC4_PACKET_CLEAR_COLORS);
cl_u32(&vc4->rcl, vc4->clear_color[0]);
cl_u32(&vc4->rcl, vc4->clear_color[1]);
cl_u32(&vc4->rcl, vc4->clear_depth);
cl_u8(&vc4->rcl, 0);
cl_start_reloc(&vc4->rcl, 1);
cl_u8(&vc4->rcl, VC4_PACKET_TILE_RENDERING_MODE_CONFIG);
cl_reloc(vc4, &vc4->rcl, ctex->bo, csurf->offset);
cl_u16(&vc4->rcl, width);
cl_u16(&vc4->rcl, height);
cl_u16(&vc4->rcl, ((csurf->tiling <<
VC4_RENDER_CONFIG_MEMORY_FORMAT_SHIFT) |
(vc4_rt_format_is_565(csurf->base.format) ?
VC4_RENDER_CONFIG_FORMAT_BGR565 :
VC4_RENDER_CONFIG_FORMAT_RGBA8888) |
VC4_RENDER_CONFIG_EARLY_Z_COVERAGE_DISABLE));
/* The tile buffer normally gets cleared when the previous tile is
* stored. If the clear values changed between frames, then the tile
* buffer has stale clear values in it, so we have to do a store in
* None mode (no writes) so that we trigger the tile buffer clear.
*/
if (vc4->cleared & PIPE_CLEAR_COLOR0) {
cl_u8(&vc4->rcl, VC4_PACKET_TILE_COORDINATES);
cl_u8(&vc4->rcl, 0);
cl_u8(&vc4->rcl, 0);
cl_u8(&vc4->rcl, VC4_PACKET_STORE_TILE_BUFFER_GENERAL);
cl_u16(&vc4->rcl, VC4_LOADSTORE_TILE_BUFFER_NONE);
cl_u32(&vc4->rcl, 0); /* no address, since we're in None mode */
}
for (int y = 0; y < ytiles; y++) {
for (int x = 0; x < xtiles; x++) {
bool end_of_frame = (x == xtiles - 1 &&
y == ytiles - 1);
/* Note that the load doesn't actually occur until the
* tile coords packet is processed.
*/
if (resolve_uncleared & PIPE_CLEAR_COLOR) {
cl_start_reloc(&vc4->rcl, 1);
cl_u8(&vc4->rcl, VC4_PACKET_LOAD_TILE_BUFFER_GENERAL);
cl_u8(&vc4->rcl,
VC4_LOADSTORE_TILE_BUFFER_COLOR |
(csurf->tiling <<
VC4_LOADSTORE_TILE_BUFFER_FORMAT_SHIFT));
cl_u8(&vc4->rcl,
vc4_rt_format_is_565(csurf->base.format) ?
VC4_LOADSTORE_TILE_BUFFER_BGR565 :
VC4_LOADSTORE_TILE_BUFFER_RGBA8888);
cl_reloc(vc4, &vc4->rcl, ctex->bo,
csurf->offset);
}
cl_u8(&vc4->rcl, VC4_PACKET_TILE_COORDINATES);
cl_u8(&vc4->rcl, x);
cl_u8(&vc4->rcl, y);
cl_start_reloc(&vc4->rcl, 1);
cl_u8(&vc4->rcl, VC4_PACKET_BRANCH_TO_SUB_LIST);
cl_reloc(vc4, &vc4->rcl, vc4->tile_alloc,
(y * xtiles + x) * 32);
if (vc4->resolve & PIPE_CLEAR_COLOR0) {
if (end_of_frame) {
cl_u8(&vc4->rcl,
VC4_PACKET_STORE_MS_TILE_BUFFER_AND_EOF);
} else {
cl_u8(&vc4->rcl,
VC4_PACKET_STORE_MS_TILE_BUFFER);
}
} else {
assert(!"unfinished: Need to end the frame\n");
}
}
}
}
